01
Problem and philosophy
Forge RC targets the gap between diagnosis and harvest. An equation can fail a physical check, or a conjecture can contain undefined symbols, yet still become an idea through aggregate scores. Viability and path completeness become independent harvest requirements.
Exploration value is separated from eligibility for harvest. Failed repairs become records, incomplete transitions become blockers, and weak priors cap verdict strength. Failures are workspace artifacts rather than passing remarks in a summary.
02
From input to candidate
EquationTermModel resolves operator, state and coefficient units. ViabilityVerdict distinguishes viable, repairable and rejected equations. The harvest firewall checks the physics-to-equation-to-theorem-to-idea path, equation viability and conjecture symbol definitions.
A mutation router maps diagnostic failures to repair operators. CoMath adds a message bus, workstream coordination and claim provenance. Package audits inspect schemas, links and JSONL synchronization, while synthetic priors cap the strongest permissible status.
03
Modules and state
Failure is itself a useful research artifact when it can be located, read and routed to repair. Candidate status remains bounded by the strength of prior evidence.
ViabilityVerdict separately records bridge, harvest and Aurora permissions, blocking reasons and repair signals. A repairable candidate may remain in the workflow without becoming harvestable; rejected artifacts can still be submitted for an audit trail.
Read component contracts and call sites separately. RC calls the harvest firewall, while repairs, units and coordination are not fully connected. This avoids describing RC as functionless or assigning Final’s integration to it.
- Unit resolution
EquationTermModel combines operator, state and coefficient units.
- Viability and firewall
ViabilityVerdict separates viable, repairable and rejected candidates; harvesting requires complete P→E→T→I links, defined symbols and viable equations.
- Repair and coordination
MutationSignalRouter routes diagnostics to repairs, while CoMath adds a persistent message bus, workstream coordination and claim provenance.
- Audit and caps
Package audits check nested schemas, ID links and JSONL synchronization; synthetic-prior strength caps the verdict.
04
Inputs, outputs and contracts
Inputs retain state and coefficient units, diagnostic findings, completed bridge legs and definitions for conjecture symbols. The firewall also reads prior strength so complete formatting cannot exceed the strength of available evidence.
Designed outputs include viability verdicts, failure records, repair traces, workstream states, messages and claim provenance. The RC implements individual components, but their presence does not prove full integration into the default inner loop; call sites matter.
- Inputs
Spark equations and bridge clusters, conjecture definitions, physics diagnostics and prior strength.
- Outputs
Structured failed-cluster/exploration reasons, viability categories, three package-audit reports and a prior-capped discovery package.
05
A conceptual reading scenario
Consider a conjecture with an incomplete theorem-to-idea transition and undefined symbols. A high aggregate score should not bypass the firewall: missing legs and symbols are listed, harvesting is blocked, and any later repair retains those failure reasons.
06
Changes from the preceding stage
Compared with Spark’s findings and suggestions, Forge RC introduces viability, repair and harvest protocols. Unit resolution explicitly includes coefficients rather than equating an operator’s dimension with the dimension of its complete term.
This is an RC snapshot. The Final audit reports incomplete inner-loop integration and honest failure on the default path. That is an unfinished workflow boundary, not evidence that autonomous agents completed research because components and tests exist.
07
What the source establishes
Source inspection confirms the firewall, unit models, routing and coordination structures. RC and Final share an internal v5.5-forge label but differ in archive and call-site integration. Keeping the RC label prevents assigning later capabilities to this snapshot.
Firewalls, the unit model, signal router, workstream/message-bus structures and audit utilities exist in code.
08
Historical scope and next steps
This is the Forge RC: the Final audit explicitly identifies incomplete live inner-loop wiring and a default honest-failure result. Only a synthetic prior backend ships; coordination records do not imply autonomous proof or literature research. RC and Final share the 5.5 version tag but differ in package and implementation, so their qualifiers must remain distinct.
Reading example · Explaining the workflow
Reading example: trace one candidate
Consider a conjecture with an incomplete theorem-to-idea transition and undefined symbols. A high aggregate score should not bypass the firewall: missing legs and symbols are listed, harvesting is blocked, and any later repair retains those failure reasons.
- 01
Read the high-scoring candidate’s completed bridge legs.
- 02
Check definitions for every conjecture symbol.
- 03
Inspect firewall reasons and equation viability.
- 04
Follow failure records to repairs and verify actual integration.
Locate the saved candidate, diagnostics and status, and identify evidence that remains absent.
This is a conceptual reading scenario derived from archive structures, not a new run, discovery result or performance experiment.
Changes in this version
Harvest eligibility
- Earlier approach
Scores may obscure failures
- This version’s approach
Independent viability and path firewall
Check eligibility before packaging an idea.
Unit model
- Earlier approach
Term dimensions omit coefficient effects
- This version’s approach
Resolve operator × state × coefficient
Identify unit-inconsistent terms.
Failure artifacts
- Earlier approach
Diagnostics remain findings
- This version’s approach
Structured failures and repair traces
Preserve failures as later inputs.